Markov Length: Scale, Memory & Transition

• Markov length is the minimal scale or memory length beyond which conditional dependence becomes negligible, serving as a local operational boundary in various fields.
• It is quantified via the exponential decay of conditional mutual information in systems ranging from mixed-state quantum matter to turbulence, indicating a transition in behavior.
• Practically, Markov length informs diagnostics in fault-tolerant quantum error correction, statistical inference in variable-memory processes, and branch determinations in phylogenetics and random media.

Markov length is not a single universal invariant. Across contemporary research, it denotes a characteristic scale or minimal memory at which conditional dependence becomes effectively local, finite, or operationally ignorable. In mixed-state quantum matter and fault-tolerant quantum error correction, it is defined through the exponential decay of conditional mutual information (CMI), including the “spacetime Markov length” extracted from repeated syndrome-measurement histories and used as a decoder-independent diagnostic of the intrinsic breakdown of fault tolerance (Negari et al., 2024). In stochastic-process theory, it denotes the relevant history or context length of a process with variable memory; in random media it denotes a mean chord length; in turbulence it denotes a Markov–Einstein coherence length; and in phylogenetics it is used as a branch length derived from the determinant of a substitution matrix (Larmier et al., 2017, Ju, 27 Apr 2026, Casanellas et al., 2011).

1. Taxonomy of the term

The common structure behind these uses is a reduction of dependence: a buffer region, a finite suffix of the past, a domain boundary, or a scale separation becomes sufficient to mediate the relevant statistics. What differs is the object being measured: entropy-based conditional dependence, memory depth, domain size, cascade coherence, or expected substitutions per site.

Domain	Meaning of “Markov length”	Representative sources
Mixed-state phases and QEC	Length scale from exponential decay of CMI	(Sang et al., 2024, Negari et al., 2024, Chen et al., 8 Dec 2025)
VLMCs and finitarily Markovian processes	Minimal relevant history or context length	(0712.0105, 0808.2964, Zambom et al., 2019)
Markov binary mixtures	Mean chord length of random domains	(Larmier et al., 2017)
Turbulence	Markov–Einstein coherence length in cascade coordinates	(Ju, 27 Apr 2026)
Phylogenetics	Branch length from $-\frac{1}{4}\log \det P$ for DNA models	(Casanellas et al., 2011)

This dispersion of meanings is substantive, not merely terminological. In the quantum-information usage, Markov length is an information-theoretic correlation scale. In the VLMC literature, it is a memory-depth concept. In random-media transport, it is a geometric parameter of disorder. In turbulence and phylogenetics, it is a modeling length attached to scale-to-scale evolution or evolutionary substitution.

2. Conditional mutual information and entropy-based definitions

In the mixed-state literature, Markov length is defined via conditional mutual information. For a tripartition $A,B,C$, the quantum CMI is

$I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$

with $S(\cdot)$ the von Neumann entropy (Sang et al., 2024). The state has Markov length $\xi$ if this quantity decays exponentially when $B$ separates $A$ from $C$,

$$I_\rho(A:C|B)\le \mathrm{poly}(|A|,|C|)\,e^{-{\rm dist}(A,C)/\xi},$$

and a family has finite Markov length when $\xi$ is independent of system size (Sang et al., 2024).

The spacetime version adapts the same logic to repeated syndrome extraction. A syndrome extraction circuit produces a spacetime lattice of measurement outcomes, partitioned into regions $A,B,C$0, $A,B,C$1, and $A,B,C$2, with $A,B,C$3 acting as a buffer. The relevant quantity is the classical CMI

$A,B,C$4

where $A,B,C$5 is the Shannon entropy of the measurement bits in the specified region (Negari et al., 2024). The spacetime Markov length $A,B,C$6 at noise rate $A,B,C$7 is extracted from

$A,B,C$8

In this formulation, finite Markov length means that conditional correlations are short ranged in spacetime, while a diverging Markov length indicates long-range conditional dependence (Negari et al., 2024).

A closely related non-equilibrium usage appears in directed percolation. There the Markov length $A,B,C$9 is defined from the exponential decay of CMI with the size of a buffer region, and the decay law

$I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$0

is used as the operational definition (Chen et al., 8 Dec 2025).

3. Mixed-state phases, local reversibility, and fault tolerance

A central development is the promotion of Markov length to a phase diagnostic. For states evolving under local Lindbladians, finite Markov length along the evolution implies that the state remains in the same mixed-state phase, in the sense that another quasi-local Lindbladian evolution can reverse it (Sang et al., 2024). The mechanism is tied to approximate recovery: the paper uses an approximate Petz recovery bound,

$I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$1

to connect small CMI to quasi-local reversibility (Sang et al., 2024). In the formulation applied to non-equilibrium systems, finite Markov length is stated to be both necessary and sufficient for local reversibility (Chen et al., 8 Dec 2025).

The fault-tolerance application is more specific. The spacetime Markov length is defined directly from classical syndrome data and is independent of the decoder; its divergence signals the intrinsic breakdown of fault tolerance (Negari et al., 2024). The underlying correspondence relates local stabilizer codes with measurement and physical errors to mixed-state phases of decohered resource states in one higher dimension (Negari et al., 2024). In the detailed formulation, the spacetime Markov length remains finite if and only if the decoding problem is physically recoverable; as the threshold is approached, the correlation length diverges, and this appears as a diverging Markov length (Negari et al., 2024).

The toric-code examples sharpen this picture. For the toric code subject to decoherence, Markov length is finite everywhere except at the decodability transition, where it diverges, and the corresponding CMI can be mapped to the free-energy cost of point defects in the random bond Ising model (Sang et al., 2024). In the spacetime fault-tolerance setting, the toric code under bit-flip noise exhibits a divergence at the critical noise rate $I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$2, matching the decoding threshold (Negari et al., 2024). The same source states that below threshold the Markov length is finite, at threshold it diverges, and above threshold it may return finite while the code lies in a trivial phase (Negari et al., 2024).

In non-equilibrium classical dynamics, the Domany–Kinzel model exhibits analogous behavior. Tensor-network simulations provide evidence for local reversibility within the active phase, while the Markov length diverges upon approaching the $I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$3-dimensional directed-percolation critical point (Chen et al., 8 Dec 2025). The same work emphasizes a contrast with classical equilibrium transitions, where Markov length is zero due to Gibbs character (Chen et al., 8 Dec 2025). For compact directed percolation, the Markov length diverges throughout the phase diagram because of spontaneous breaking of domain-wall parity symmetry from strong to weak, yet the conditional mutual information continues to detect the phase transition (Chen et al., 8 Dec 2025).

4. Variable memory, context trees, and statistical inference

In stochastic-process theory, Markov length is a memory-depth notion. For finitarily Markovian processes, the memory length is the least finite $I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$4 such that the conditional distribution of the next symbol given the entire past equals the conditional distribution given only the last $I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$5 symbols (0712.0105). A related formulation defines a memory word as a finite word $I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$6 for which the conditional probability of the present value is constant on the cylinder set defined by that word, and a minimal memory word as one with no proper suffix having the same property (0808.2964).

This perspective generalizes fixed-order Markov chains into variable-length Markov chains (VLMCs). In a VLMC, the relevant past is a variable-length suffix determined by a context tree rather than a fixed order (Cénac et al., 2010). The formal memory structure can be encoded through the longest internal suffix and the associated $I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$7-lis decomposition, which control recursive computation of cylinder probabilities and characterize existence and uniqueness of stationary probability measures (Cénac et al., 2018). In another treatment, the pair $I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$8, where $I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC),$9 records the current memory, is Markovian even when the observed letter process $S(\cdot)$0 is not (Cénac et al., 2012).

The inference problem is correspondingly about estimating the relevant history length. One line of work gives a universal estimator that converges almost surely to the length of the longest minimal memory word, while proving that no such universal estimator exists for the length of the shortest memory word (0808.2964). Another line gives a rather complete analysis for finitarily Markovian processes: backward estimation of memory length is universally consistent, whereas forward estimation is possible only along stopping times with density arbitrarily close to one, and universal estimation with density one is impossible (0712.0105).

When exogenous covariates are present, the relevant Markov length becomes context-specific. The “beta-context algorithm” estimates a maximal context tree, fits context-dependent logistic regressions, and then prunes lags and covariates by likelihood-ratio tests; the resulting estimator of the context tree and parameters is strongly consistent (Zambom et al., 2019). Under stochastic contamination, a two-step method first estimates a complete order-$S(\cdot)$1 hidden Markov chain via an adapted Baum–Welch procedure and then recovers the variable-length dependence structure by a bootstrap Bayesian Information Criterion, again with consistency guarantees (Duarte et al., 2017).

5. Geometric, transport, turbulence, and evolutionary uses

In particle transport through random media, Markov length means mean chord length. For a $S(\cdot)$2-dimensional Poisson tessellation, the mean chord length $S(\cdot)$3 is the average length of a straight segment traversed in a random domain before hitting an interface, with

$S(\cdot)$4

where $S(\cdot)$5 is the Poisson tessellation density (Larmier et al., 2017). In an infinite medium the chord-length distribution is exponential,

$S(\cdot)$6

and this law is used in Chord Length Sampling to generate interface crossings on the fly (Larmier et al., 2017). In this setting, the Markov length is also a correlation length for the spatial disorder: small $S(\cdot)$7 corresponds to fine mixing, and the approximation improves as $S(\cdot)$8 decreases and as dimension increases (Larmier et al., 2017).

In turbulence, the relevant notion is the Markov–Einstein coherence length $S(\cdot)$9 in log-scale cascade coordinates. Using direct numerical simulation at $\xi$0 and $\xi$1, together with two Markov-by-construction null surrogates, the turbulent cascade was measured to have $\xi$2–$\xi$3 for the full cascade, approximately three times the canonical estimate $\xi$4 (Ju, 27 Apr 2026). The scale coordinate is

$\xi$5

and a gap-scan based on the Wilcoxon rank-sum test identifies the smallest separation at which excess rejection falls to the surrogate baseline (Ju, 27 Apr 2026). Stratified analyses show that intermittent events carry $\xi$6–$\xi$7, while at mid-inertial-range scales the quiescent cascade recovers $\xi$8–$\xi$9; near the dissipation range the pattern reverses, consistent with the spectral bottleneck (Ju, 27 Apr 2026).

In phylogenetics, Markov length is used as branch length, namely the expected number of substitutions per site or a determinant-based surrogate. For DNA models, the standard expression is

$B$0

where $B$1 is the transition matrix (Casanellas et al., 2011). In continuous-time models with $B$2, this matches the expected substitutions per site through the trace of the rate matrix; in more general discrete-time models it remains the determinant-based quantity used to prescribe branch length when constructing substitution matrices under models such as JC*, K80*, K81*, SSM, and GMM (Casanellas et al., 2011).

6. Terminological distinctions and conceptual unification

Because the phrase is overloaded, several distinctions are essential. First, Markov length is not generally the same as chain length. In the toric homogeneous Markov chain literature, the paper explicitly does not use the phrase “Markov length”; instead it studies the chain length $B$3, the number of time points, which controls the contingency-table structure, the fiber, and the complexity of the Markov basis for conditional tests (Takemura et al., 2010). Second, it is not always a correlation length. In Markov stick-breaking processes, “Markov length variables” are the stick-breaking variables $B$4 generated as a Markov process on $B$5, rather than a spatial or temporal scale extracted from decay (Gil-Leyva et al., 23 Jan 2026).

The most coherent unifying statement supported by the literature is therefore structural rather than definitional. In quantum phases and spacetime error correction, Markov length is the scale at which a buffer makes distant regions conditionally independent, and its divergence marks a transition or the breakdown of fault tolerance (Sang et al., 2024, Negari et al., 2024). In variable-memory processes, it is the minimal suffix required to retain predictive sufficiency (0712.0105, 0808.2964). In random media and turbulence, it is the operational scale over which spatial or cascade correlations persist (Larmier et al., 2017, Ju, 27 Apr 2026). In phylogenetics, it is the length assigned to a stochastic transformation through $B$6 (Casanellas et al., 2011).

A common misconception is to treat these objects as interchangeable because they share the word “Markov.” The sources do not support that identification. They support, instead, a family resemblance: Markov length names the point at which a Markovian approximation, a finite-memory truncation, or a local conditional description becomes valid—or ceases to be valid.